Vision-Language Models are Zero-Shot Reward Models for Reinforcement Learning
Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel pith:2MNSEPXWrecord.jsonopen to challenge →
read the original abstract
Reinforcement learning (RL) requires either manually specifying a reward function, which is often infeasible, or learning a reward model from a large amount of human feedback, which is often very expensive. We study a more sample-efficient alternative: using pretrained vision-language models (VLMs) as zero-shot reward models (RMs) to specify tasks via natural language. We propose a natural and general approach to using VLMs as reward models, which we call VLM-RMs. We use VLM-RMs based on CLIP to train a MuJoCo humanoid to learn complex tasks without a manually specified reward function, such as kneeling, doing the splits, and sitting in a lotus position. For each of these tasks, we only provide a single sentence text prompt describing the desired task with minimal prompt engineering. We provide videos of the trained agents at: https://sites.google.com/view/vlm-rm. We can improve performance by providing a second "baseline" prompt and projecting out parts of the CLIP embedding space irrelevant to distinguish between goal and baseline. Further, we find a strong scaling effect for VLM-RMs: larger VLMs trained with more compute and data are better reward models. The failure modes of VLM-RMs we encountered are all related to known capability limitations of current VLMs, such as limited spatial reasoning ability or visually unrealistic environments that are far off-distribution for the VLM. We find that VLM-RMs are remarkably robust as long as the VLM is large enough. This suggests that future VLMs will become more and more useful reward models for a wide range of RL applications.
This paper has not been read by Pith yet.
Forward citations
Cited by 9 Pith papers
-
Improving Robotic Generalist Policies via Flow Reversal Steering
Flow Reversal Steering steers flow matching generalist policies by reversing suboptimal actions to nearby better modes, enabling improved zero-shot control, quick distillation, and RL bootstrapping in robotic manipulation.
-
World Model Self-Distillation: Training World Models to Solve General Tasks
Self-distillation from a caption-conditioned video diffusion model to an image-and-prompt-conditioned executor, enhanced by RL from VLM feedback, enables task solving in world models.
-
VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models
VoxPoser uses LLMs to compose 3D value maps via VLM interaction for model-based synthesis of robust robot trajectories on open-set language-specified manipulation tasks.
-
Learning Process Rewards via Success Visitation Matching for Efficient RL
Success Visitation Matching uses a discriminator to turn sparse outcome rewards into dense process rewards by matching visitations of successful episodes, provably preserving the optimal policy and speeding up robotic...
-
OrganicHAR: Towards Activity Discovery in Organic Settings for Privacy Preserving Sensors Using Efficient Video Analysis
OrganicHAR discovers 4-8 activity categories per user from sensor signals, achieves 79% accuracy on coarse activities with ambient sensors alone and cuts VLM queries by 90% by triggering video analysis only at detecte...
-
A Survey on Generative Recommendation: Data, Model, and Tasks
This survey organizes generative recommendation into data, model, and task dimensions, identifying five advantages including world knowledge integration and creative generation while noting challenges in benchmarks an...
-
LLM-as-a-Verifier: A General-Purpose Verification Framework
Computing the expectation over scoring-token logits instead of taking argmax enables verification to scale along granularity, repetition, and criteria decomposition, achieving state-of-the-art on four agentic benchmar...
-
AtlasVA: Self-Evolving Visual Skill Memory for Teacher-Free VLM Agents
AtlasVA organizes VLM agent memory into spatial heatmaps, visual exemplars, and symbolic skills, evolving atlases from trajectories to act as potential-based shaping rewards in teacher-free reinforcement learning.
-
Vision-Language Models for Deployable Social Robot Navigation: Bridging Semantic Reasoning and Low-Level Control
Survey organizing VLM-based social robot navigation into reasoning, planning, and bridging components with a proposed roadmap for hybrid deployable systems.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.